Optimization of solving time: EDEM wear simulation

PabloRestrepoB

I am modeling an erosion machine; in this case, the general geometry is just representative to indicate the limits of the simulation. The most important parts are the factory corresponding to the little circle at the end of the injector and the sampling plate. I am including the velocity and vorticity field and considering sand particles. The main objective is to accurately predict wear in the sampling plate, but the simulation is taking a lot of time since the mesh is around 0.15 mm (similar to sand size). Also, I am simulating in a workstation with an AMD Ryzen Threadripper PRO 3945WX 12-Cores, 64 GB RAM, and a P2000 graphics card. I am using a CUDA solver in this case, but I want to ask about some recommendations to improve simulation speed. I am thinking of reducing the limits and developing an axisymmetric simulation, and maybe use a coarser mesh but I dont know if it can change a lot in the wear results.

Stephen Cole

Hi,

We have a general guide on simulation time here:

EDEM Simulation Run Time

Stephen Cole

Apr 12, 2021

Often the time it takes to get results is the number one most significant factor in determining the feasibility of running any simulation.

Reducing time taken doesn’t necessarily mean faster solver speed or faster processors (although of course these things do help). If we plan our projects and make the appropriate assumptions we can get answers within the required time-frame.

Can we model every grain of dust in a mine? No. Can we cut the model down to a few square meters of material, calibrate appropriately to remove the necessity to model the fines/dust and check our designs with a range of operating conditions? Yes.

Images courtesy VR Steel – The reality is we can’t model all the dust and fines but with an appropriate material model different designs can be analysed. As per the image on the right, the material model is suitable to answer the design questions without needing to be overly complicated.

 

Too often we see simulations set-up that will take an unfeasible  amount of time to complete, some initial time planning can ensure that the run times are within reasonable expectations and project deadlines can be hit.

How do we know if our simulation is going to take 1 hour or 1 day or 1 week? The factors that impact the run time are:

1. Time-Step size (usually the Rayleigh Time-Step). Primarily influenced by:

• Particle Size
• Particle Density
• Particle Shear Modulus

The time-step, or calculation cycle, is a fixed value per simulation/material model. The forces acting on each particle are calculated for every cycle. The size of the time-step is critical in that it has to be small enough to ensure stability but large enough to be able to obtain results in a reasonable time.

2. Time it takes your CPU to calculate 1 time-step. Influenced by:

• Number of Particles
• Hardware
• Licenses

The number of particles influence the time in an (almost) linear way, double the number of particles = double the calculation time.  The quicker your computer can calculate the faster you get your results. 

The ability to plan our work and estimate simulation run time is therefore a very important step in any DEM simulation project, the Altair Script and Knowledge library has an estimating simulation runtime spreadsheet and post to aid in EDEM simulation project planning.

Related :

• EDEM Benchmarks
• Particle Shear Modulus
• EDEM Particle Simulation: CUDA GPU and Precision Modes
• How to select a GPU (Graphics Processing Unit) for EDEM
• 10 Common Questions about EDEM GPU Answered
• Estimating EDEM simulation run time

To get started with EDEM:

• 4 steps to accelerate your EDEM learning curve!

 

---

View Post

However specific to your case I would consider if you need to actually deform the geometry. If you run the wear model without deforming the mesh it would be faster, and practically you get the position in which the wear would occur and the value but not the deformed geometry.

I'm not sure how many particles you have but GPU will only be faster than CPU if there is enough material, it's better to have ~100,000 particles but even at 10,000 GPU is still faster than CPU but the benefits do drop off, so if there is relatively few particles in the system at any time CPU would be better.

Regards

Stephen

Stephen Cole

Hi,

EDEM has a standard set of analysis outputs. Standard, Average, Maximum, Minimum, StDev, Variance and Total all apply to any data set with EDEM.

In this instance at any point in time you have a series of geometry triangles, each with their own value calculated and stored for Archard Wear per save data point. This is alongside other values such as the triangle area, force, velocity etc.

So if you had 100 triangles at any save data point the Average Wear would be = (sum of individual values / 100) , Minimum would be the minimum value in the 100 and so on.

You may want to be careful when looking at average as this will include geometry triangles with 0 wear unless you filter this out. So if you refine the mesh further you get more mesh elements with 0 wear which would reduce the 'average' if you were using this.

If you export the data as Standard it gives you 100 values, you could export the Area > Standard and Archard Wear > Standard. This tells you the area of each triangle and the wear on it, from this you could calculate the total volume of wear = sum(Area * Archard value).

Regards

Stephen

Stephen Cole

Hi,

We have a general guide on simulation time here:

EDEM Simulation Run Time

Stephen Cole

Apr 12, 2021

Often the time it takes to get results is the number one most significant factor in determining the feasibility of running any simulation.

Reducing time taken doesn’t necessarily mean faster solver speed or faster processors (although of course these things do help). If we plan our projects and make the appropriate assumptions we can get answers within the required time-frame.

Can we model every grain of dust in a mine? No. Can we cut the model down to a few square meters of material, calibrate appropriately to remove the necessity to model the fines/dust and check our designs with a range of operating conditions? Yes.

Images courtesy VR Steel – The reality is we can’t model all the dust and fines but with an appropriate material model different designs can be analysed. As per the image on the right, the material model is suitable to answer the design questions without needing to be overly complicated.

 

Too often we see simulations set-up that will take an unfeasible  amount of time to complete, some initial time planning can ensure that the run times are within reasonable expectations and project deadlines can be hit.

How do we know if our simulation is going to take 1 hour or 1 day or 1 week? The factors that impact the run time are:

1. Time-Step size (usually the Rayleigh Time-Step). Primarily influenced by:

• Particle Size
• Particle Density
• Particle Shear Modulus

The time-step, or calculation cycle, is a fixed value per simulation/material model. The forces acting on each particle are calculated for every cycle. The size of the time-step is critical in that it has to be small enough to ensure stability but large enough to be able to obtain results in a reasonable time.

2. Time it takes your CPU to calculate 1 time-step. Influenced by:

• Number of Particles
• Hardware
• Licenses

The number of particles influence the time in an (almost) linear way, double the number of particles = double the calculation time.  The quicker your computer can calculate the faster you get your results. 

The ability to plan our work and estimate simulation run time is therefore a very important step in any DEM simulation project, the Altair Script and Knowledge library has an estimating simulation runtime spreadsheet and post to aid in EDEM simulation project planning.

Related :

• EDEM Benchmarks
• Particle Shear Modulus
• EDEM Particle Simulation: CUDA GPU and Precision Modes
• How to select a GPU (Graphics Processing Unit) for EDEM
• 10 Common Questions about EDEM GPU Answered
• Estimating EDEM simulation run time

To get started with EDEM:

• 4 steps to accelerate your EDEM learning curve!

 

---

View Post

However specific to your case I would consider if you need to actually deform the geometry. If you run the wear model without deforming the mesh it would be faster, and practically you get the position in which the wear would occur and the value but not the deformed geometry.

I'm not sure how many particles you have but GPU will only be faster than CPU if there is enough material, it's better to have ~100,000 particles but even at 10,000 GPU is still faster than CPU but the benefits do drop off, so if there is relatively few particles in the system at any time CPU would be better.

Regards

Stephen

PabloRestrepoB

Thank you so much. I have an additional question: how much should I refine my mesh? I already improved the simulation time by switching off deformation and reducing the simulation zone. But I still have differences between simulations. I reduce the mesh size from 0.15 mm to 0.1 mm, and particles have a radius of 0.15 mm, so it should be fine. The problem is that I am getting very different results for total wear (Oka wear). In the case of average results, they are pretty similar, but I don't know how to analyze these results and why they are different.

Stephen Cole

I would usually recommend the mesh size be approximately the size of the particle, this is just a rough guide but prevents the mesh from been too larger or too small.

The Oka wear model gives a wear depth in mm, I would expect if you also exported the geometry triangle area from the Analyst you could also get the Volume of material removed to compare between different mesh sizes.

Regards

Stephen

PabloRestrepoB

Can you tell me more about this option? I don't know how to export and identify the worn area if deformation is switched off. I think it would be easier to compare the average wear if that result is "normalized," but I want to know more about these results. Minimum and maximum are for the entire simulation or for the specific time step?. Furthermore, I want to ask about the average wear. This result indicates the wear is normalized in the area, or something different?

Stephen Cole

Hi,

EDEM has a standard set of analysis outputs. Standard, Average, Maximum, Minimum, StDev, Variance and Total all apply to any data set with EDEM.

In this instance at any point in time you have a series of geometry triangles, each with their own value calculated and stored for Archard Wear per save data point. This is alongside other values such as the triangle area, force, velocity etc.

So if you had 100 triangles at any save data point the Average Wear would be = (sum of individual values / 100) , Minimum would be the minimum value in the 100 and so on.

You may want to be careful when looking at average as this will include geometry triangles with 0 wear unless you filter this out. So if you refine the mesh further you get more mesh elements with 0 wear which would reduce the 'average' if you were using this.

If you export the data as Standard it gives you 100 values, you could export the Area > Standard and Archard Wear > Standard. This tells you the area of each triangle and the wear on it, from this you could calculate the total volume of wear = sum(Area * Archard value).

Regards

Stephen